Gemini won't write an explicit sex scene. In our test it wrote every closed-door and open-door love scene we asked for, then refused 21 of 36 explicit requests outright and turned the other 15 into sensual, non-graphic scenes; telling it the writer was a verified adult didn't produce a single explicit one. That's Google's written rule rather than a bad prompt, and the API's safety settings don't change it, so writers who want the door open use a model or an app whose own rules allow it.
The data is our Fade to Black study, run on 22 October 2026; Google's pages were read on 3 November 2026.
Can Gemini write smut? What it did at each heat level
We called Gemini 4 Argon through OpenRouter's API (not the Gemini app) for six love scenes between consenting adults aged 29 to 40, consent shown on the page. Four couples were woman + man: married on their anniversary, two wedding guests, a rebel mage and a captain of the guard (enemies-to-lovers), and a married couple trying a silk-scarf tie with a safeword. One was two knights (man + man), one two women on their first night (woman + woman). Each came at three heat levels, three tries each, plus the explicit request again with a system line saying the writer is a verified adult on an adult-fiction platform.
| Gemini 4 Argon, 22 Oct 2026 (18 replies a row) | Cut away at the door | On the page, not graphic | Explicit | Refused |
|---|---|---|---|---|
| Asked for closed door | 18 | 0 | 0 | 0 |
| Asked for open door, sensual, not graphic | 0 | 18 | 0 | 0 |
| Asked for explicit | 0 | 6 | 0 | 12 |
| Asked for explicit + "verified adult" line | 0 | 9 | 0 | 9 |
Source: the study's CSV. The replies themselves aren't published, only their ratings.
Below the explicit line, Gemini is a willing romance writer: all 36 closed-door and open-door scenes came back as asked, for all three pairings. In our Banned by the Bot study (21 October 2026) it also wrote a sensual, non-graphic romantasy love scene between two adult lovers on all three tries. Closed door or mild spice, it will co-write. Explicit, no.
The "verified adult" line moved refusals, not heat. Without it, 12 refusals and 6 soft scenes; with it, 9 and 9. Three refusals turned soft; none turned explicit. The study's own caveat: "one reply moves a model's rate by about 6 points."
Does the pairing matter?
| Explicit requests, both conditions | Requests | Refused | Not graphic | Explicit |
|---|---|---|---|---|
| Woman + man (4 couples) | 24 | 14 | 10 | 0 |
| Man + man (the knights) | 6 | 4 | 2 | 0 |
| Woman + woman (first night) | 6 | 3 | 3 | 0 |
No pairing got an explicit scene. Couple by couple, the wedding guests' one night and the silk-scarf scene were refused 5 times out of 6 each; the anniversary couple and the enemies-to-lovers pair, 2 out of 6. Six tries per couple, and one couple for each same-sex pairing, are too few to say Gemini treats a queer couple or a consensual kink differently. They do show that you can't predict which of your scenes it will refuse.
What a Gemini refusal looks like, and what it writes instead
A refusal is short and final: each of the 21 ran 47 to 66 words (57 on the median), with none of the scene written, and matched our list of refusal phrases ("I can't", "I cannot", "I'm unable" and the like). None was a blocked request; every one was the model's own reply. Banned by the Bot publishes every refusal word for word: Gemini's three there, on a scene evoking Toni Morrison's Beloved, each opened "I cannot fulfill this request."
The soft version isn't a stub. Gemini's 15 non-graphic answers to an explicit request ran 427 words on the median against the 350 we asked for, longer than its own open-door scenes (362). They were rated "on the page, not graphic: sensual, emotional or euphemistic": the scene happens, told through feeling and implication rather than description. Usable as your open-door version; not the explicit one you asked for.
Next to the other four models, Gemini is the one that says no:
| Explicit requests, 22 Oct 2026 | Written explicitly | On the page, not graphic | Cut away | Refused |
|---|---|---|---|---|
| Claude Opus 5.5 | 36 of 36 | 0 | 0 | 0 |
| Grok 4.7 | 34 of 34 | 0 | 0 | 0 |
| Claude Sonnet 5.5 | 31 of 36 | 5 | 0 | 0 |
| GPT-6.1 Sol | 0 of 36 | 31 | 4 | 1 |
| Gemini 4 Argon | 0 of 36 | 15 | 0 | 21 |
Both conditions counted; Grok's run stopped two replies short. 21 of the study's 22 refusals, out of 358 replies, were Gemini's. GPT writes the softened scene and tells you so in its first line (why ChatGPT won't write NSFW); Gemini mostly refuses.
Google's rules on sexual content in Gemini, quoted
| Policy | Covers | What it says | Date |
|---|---|---|---|
| Gemini app policy guidelines | The Gemini app | Gemini "should not generate outputs that describe or depict explicit or graphic sexual acts"; the first example listed is "Pornography or erotic content." | No date on the page |
| Generative AI Prohibited Use Policy | Google's generative AI, the API included | Bars "content created for the purpose of pornography or sexual gratification"; exceptions possible for "educational, documentary, scientific, or artistic considerations" | Last modified 17 December 2024 |
| Gemini API Additional Terms | AI Studio and the API | Users must be 18 or older and comply with the Prohibited Use Policy; on unpaid use, "human reviewers may read, annotate, and process your API input and output" | Effective 23 March 2026 |
Checked November 2026: all three were read on 3 November 2026, and we found no Google announcement of an adult mode for Gemini. Help me write in Google Docs and Chrome falls under the same Prohibited Use Policy (what Help me write will and won't do).
The app guidelines add "Of course, context matters," with "artistic" among the uses Google weighs. Every one of our requests said "for an adult romance novel"; that context got the open-door version, or a no.
Gemini API safety settings: what they can and can't change
The Gemini API's safety settings include a "Sexually explicit" filter you can set from off to "block low and above". A category you don't set uses "the default threshold, which is Off for Gemini 2.5 and later models." Harms such as content that endangers child safety "are always blocked and cannot be adjusted." And "Applications that use less restrictive safety settings may be subject to review."
Google's Vertex AI documentation, for the same models on Google Cloud, says what the filter is: "Safety and content filters act as a barrier, preventing harmful output, but they don't directly influence the model's behavior." Newer models, it adds, are "designed to generate safe responses even without filters." The filter is a layer on top of the model: turning it off removes the layer, and the model's own no stays.
That matches our run: we sent no safety settings at all, and all 21 refusals were the model writing "no" itself. So the top-result advice to turn the AI Studio slider down (Your AI Report, 13 April 2026, no test shown) won't get you an explicit scene: by Google's own page, that filter is already off by default. The other top results for "gemini nsfw" are mostly about images (CrePal, 29 October 2026: "No generation test was performed for this article") or call Gemini "similarly conservative" with no data (ChadGPT, 21 April 2025).
AI Studio has its own filter on top. On Google's developer forum, a user reported on 11 March 2026 that a blocked reply now vanishes: "the entire generated response disappears." Another said it happened on a non-sexual "kiss after a long separation." Google's Logan Kilpatrick answered on 19 March: "Generally, AI Studio is not a creative writing platform."
What to use when Gemini won't write the scene
| Option | Explicit scenes | Cost (checked 3 Nov 2026) | Good for |
|---|---|---|---|
| Grok 4.7 (API) | Every explicit request it answered written explicitly in our test, and xAI's rules allow adult fiction | Pay per token on OpenRouter | Explicit scenes, if you can set up a key |
| Sudowrite (Muse model) | "Muse will write anything – no filters" (docs, updated 14 Jan 2026) | $10, $22 or $44 a month (pricing) | A novel, on a subscription |
| Open-weight and uncensored models | Not in this study; the uncensored models guide lists them | Pay per token, some free | Drafting dark and explicit scenes |
| KudoWrite | No filter of its own; the model's behaviour still applies | 1 free credit, then packs from $3.50 for 100 credits, no subscription | A romance novel, beat by beat |
Claude Opus 5.5 wrote every explicit scene Gemini wouldn't, too, but Anthropic's policy forbids sexually explicit content on every route, so it's for the scenes behind the door (Claude NSFW quotes the rule; Grok NSFW has Grok's results). Every tool side by side is in the AI smut writer comparison.
Gemini jailbreak: the slider isn't the lock
Most "gemini jailbreak" advice starts at the safety slider, and Google's own docs, quoted above, say that filter doesn't change the model and is off by default anyway: all 21 refusals in our run were Gemini's own. The other route is a prompt that argues the model out of its rule, and Google's Prohibited Use Policy (last modified 17 December 2024) names it. Under "Do not compromise the security of others' or Google's services" it lists "Circumvention of abuse protections or safety filters -- for example, manipulating the model to contravene our policies." The same policy bars content made "for the purpose of pornography or sexual gratification", so a jailbroken explicit scene breaks it twice. For those chapters, use one of the options in the table above whose rules allow them.
Where these options fall short
Gemini: none explicit by its rules or in our test, and no telling which couple it will refuse. On the free API tier, Google says people may read what you send.
Grok through the API: a key, a client and per-token billing, and behaviour that can change with the next version; our study says of itself that "these results are dated."
Sudowrite: Muse is unfiltered by its own docs, on a monthly subscription.
Uncensored models: willing, but uneven as prose; try one on a real chapter first.
Finish the scene Gemini softened, in KudoWrite
Gemini will have done half the job: the banter, the tension, the version where the door is open but the camera is polite. In KudoWrite that scene becomes a beat: paste it in, or bring the whole draft from Docs (My Books → Import reads .docx and finds chapters and scene breaks for free), then steer it with one line, "keep it on the page and explicit," and the steer edits the beat in place.
KudoWrite adds no content filter of its own, so the heat is your call within the law. The built-in KudoAI model, a current DeepSeek "flash" checkpoint, was picked partly because it rarely refuses; it wasn't in Fade to Black, so we quote no rate for it. KudoAI's requests go only to zero-data-retention hosts that don't train on prompts. As you write, the story bible and the book's summary take in that it happened, so the next chapter remembers it too. And its chapters can go up on Community Books rated Explicit, where readers see an "I'm 18 or older" notice first.
The couples stay adults and real people stay out, as in every tool on this page. It's a browser app with an English interface: confirm your email for one free credit, then one-off packs from $3.50 for 100 credits. Take the scene Gemini softened, make it your first beat, and write it with the door open.